A Specialized Lightweight Fire Detection Model for Real-Time IoT and Embedded Applications
This repository contains the implementation of FireNet-Lite, a highly efficient CNN-based model designed for real-time fire image classification using depthwise separable convolutions. With only 7,693 trainable parameters, it is well-suited for deployment on resource-constrained devices like IoT surveillance systems and drones.
(Preprint of the research paper on this work will be available soon. Please consider citing this repository if you happen to use the code or model in your work.)
@misc{ghous2024firenetlite, title={FireNet-Lite: A Separable Convolutional Network for Ultra-Efficient Fire Image Classification}, author={Ali, Ghous and Ansari, Mohammad Samar and Ahmed, Muhammad and Yasir, Muhammad Sanad}, year={2024}, note={GitHub Repository: https://github.com/ghous-ali/FireNet-Lite} }
In our paper, we used a training dataset accessed directly from Google Drive: 🔗 Google Drive Fire Detection Dataset
for Testing 🔗 https://drive.google.com/drive/folders/1seop28o10RJANG4QCr10bYNHhvwBFInZ?usp=sharing
All training was performed using this dataset, which contains labeled images and videos for fire and non-fire scenes. Testing was conducted on held-out samples from this dataset to ensure fair and unbiased evaluation across diverse fire scenarios.
The following is the link to our Kaggle dataset used for training and evaluation:
🔗 https://www.kaggle.com/datasets/atulyakumar98/test-dataset
- ✅ Only 7,693 trainable parameters
- ✅ Accuracy: 93.00%
- ✅ Recall: 96.57%
- ✅ Precision: 91.21%
- ✅ Optimized for edge and real-time applications
For details about the architecture, evaluation metrics, and deployment scenarios, please refer to the full notebook or our upcoming paper.